Researchers affiliated with ETH Zurich’s Robotic Systems Lab describe a generalist humanoid locomotion controller in an October 22 preprint. They report real-world demonstrations on a Unitree G1 including ladder and stair climbing, crawling, vaulting, stepping-stone traversal and recovery.
A controller that can handle several kinds of physical contact could reduce the need to build a separate policy for every obstacle. But moving reliably through a complex environment also requires navigation and other capabilities that this study does not claim to solve.
What happened?
Researchers affiliated with ETH Zurich’s Robotic Systems Lab describe a generalist humanoid locomotion controller in an October 22 preprint. They report real-world demonstrations on a Unitree G1 including ladder and stair climbing, crawling, vaulting, stepping-stone traversal and recovery.
The approach combines contact planning, motion priors, reinforcement-learning fine-tuning and onboard mapping. The authors also present quantitative comparisons, primarily in simulation; those results should not be mistaken for a broad independent real-world benchmark.
The paper explicitly limits the work to local goal-conditioned locomotion. Some terrains use manually specified intermediate waypoints, and global navigation is outside its scope. The demonstrations and findings are the researchers’ reports, not an independently replicated LessCarbs test.
Source: Research authors ↗
Why does it matter?
A controller that can handle several kinds of physical contact could reduce the need to build a separate policy for every obstacle. But moving reliably through a complex environment also requires navigation and other capabilities that this study does not claim to solve.
Read the sources
- Humanoid Mobility via Attention-based Contacts ↗
Research authors · October 22, 2026
